最小化图像描述子敏感度的大规模图像索引及检索方法

With the rapid development of smart phones and other digital equipments, the number of images being stored on the Internet grows exponentially. Large scale image retrieval is to find a group of relevant or similar images for a given query image. This is vital to the utilization of this huge volume of image resources on the Internet, e.g. goods finding in e-retailing, people finding and other information retrieval problems. Owing to the large volume of images, indexing based on hashing is required to reduce the time for query. The set of images with the same or similar hash code (e.g. hamming distance = 1) of the query image's is regarded as relevant images. Then, top few relevant images ranked by a similarity measure are returned to user. We propose to study both indexing and retrieval in a unified framework of minimization of descriptor sensitivity. Changes in images, e.g. color difference, will change the values of image descriptor and will affect the quantization of hashing and similarity measure. We define these effects as the descriptor sensitivity. Current methods ignore the effects of small perturbations to image descriptors with respect to the indexing and similarity results. We define these effects as the sensitivity of indexing and similarity, respectively. Minimization of these sensitivities enhances robustness and increase precision of image indexing and retrieval. In this work, we research a hashing method via a minimization of descriptor sensitivity respect to quantization error. Then, this hashing method is extended to a hierarchical hashing method to provide different similarity among image sub-sets with the same hamming distance between their hash codes and that of the query image's. Then, a similarity measure learning method based on minimization of descriptor sensitivity is proposed to enhance the initial sub-set of images returned in query. The retrieval performance will be further improved by a sensitivity minimization based learning method with user feedback. Finally, we will integrate all components into a prototype system.

随着智能手机等摄像设备的迅速普及，互联网上的图像数量以几何级数增长。对海量图像进行大规模图像检索是网上商业活动、日常查询或医疗、交通等各行业信息检索中极为重要的应用。海量图像首先须通过哈希来建立索引以降低检索时间，与查询图像哈希编码相似的便会被看成相似的图像子集，然后使用相似度度量对该子集排序后相似度最高的少部分图像便会被返回给用户。图像的色差等变化会改变图像描述子的值并影响哈希及相似度，我们定义这个影响为敏感度。本项目提出一个最小化描述子特征敏感度框架来同时研究索引和检索问题。首先研究最小化量化误差敏感度的哈希算法以增加相似图像被划分到相同哈希码的机会，然后推广成层次哈希为哈希码差一位的图像提供不同的相似度，再研究最小化敏感度的相似度度量学习算法提升检索准确率。对哈希后的图像库学习相似度来优化初始检索效果，然后基于用户反馈和敏感度学习进一步提升检索准确率。最后整合研究结果为一个原型系统。

.大数据环境中每时每刻都有新的数据被产生，高效的检索和索引方法是有效使用大数据的关键科学问题，可是现有的方法都是以给定的固定数据库为前提，而非针对大数据中数据被不停产生并加入数据库的环境。其中语义图像检索是大数据中一个关键应用问题，所以本项目的主要研究内容便是如何在大数据动态环境下有效地进行语义图像检索。本项目首次提出大数据动态环境下的语义图像检索问题并提出增量哈希方法来解决它，成果已发表在SCI一区的IEEE T-Cyb上，这个研究成果开创了一个新的动态语义图像检索领域，打破了传统只针对给定数据库的限制，使得基于哈希的图像检索方法更为适用于现实的大数据动态环境中，对本领域的研究及应用均有重大影响。另外本项目也提出通过使用不对称哈希来使用长哈希码进行检索及短哈希码进行存储以同时提高检索及存储效率，并于SCI二区的IEEE TMM上发表并获得比在发表在其它高水平期刊上的方法更优的检索效果。而本项目的另一成果两层哈希方法先把图像描述向量投映到高维二元空间再过通最小误差方法投映到低维的二元空间获得短的哈希码，有助减少因为直接投从图像描述向量的实数空间投映到低维二元哈希码的大量信息掉失，从而获得更优的哈希检索效果，该成果已于SCI二区的Neurocomputing上发表。项目已发表9篇高水平科研论文，包括2篇SCI一区、5篇SCI二区文章，另有部分文章在审稿中。综合而言，本项目的科研产出水平较高，不但开发出比现有方法更好的哈希算法，更开创了大数据动态哈希这个前的研究方向，成果义意重大。